| Abstrakt: |
Background: One of the best ways to reduce the spread of tuberculosis (TB) is to diagnose the disease using chest X-ray (CXR) images as a low-cost and affordable method. However, there are two problems: the lack of adequate radiologists and the possibility of misdiagnosis. This is why it is necessary to use an accessible and accurate diagnostic system. This research aimed to design an accurate and accessible automatic diagnosis system that can solve diagnosis problems using deep learning. Methods: Six convolutional neural networks (CNNs), InceptionV3, ResNet50, DenseNet201, MnasNet, MobileNetV3, and EfficientNet-B4, were trained by transfer learning, the Adam optimizer, and 20 training epochs using the new, large, and accurate TBX11K dataset. The network was designed to categorize images into three groups: patients diagnosed with TB, patients exhibiting lung abnormalities unrelated to TB, and healthy individuals with no evidence of TB or other pulmonary anomalies within the lung imagery. Results: In the testing step, the networks achieved very high performance. The EfficientNet-B4 network outperformed the other networks with a sensitivity of 97.1%, specificity of 99.9%, and accuracy of 99.5%. It also performed better than previous studies in TB diagnosis using CXR images by CNNs. Conclusion: This research showed that with access to large high-quality datasets and standard training, it is possible to entrust the diagnosis of TB using medical images to computers and artificial neural networks with high confidence as they achieved accuracies higher than 99%. [ABSTRACT FROM AUTHOR] |
